Hybrid Method for the Detection of Evasion Attacks Aimed at Machine Learning Systems

Abstract

The existing methods for the detection of evasion attacks in machine learning systems are analyzed. An experimental comparison of the methods is carried out. The uncertainty method is universal; however, in this method, it is difficult to determine such uncertainty boundaries for adversarial examples that would enable the precise identification of evasion attacks, which would result in lower efficiency parameters with respect to the skip gradient method (SGM) attack, maps of significance (MS) attack, and boundary attack (BA) compared to the other methods. A new hybrid method representing the two-stage input data verification complemented with preliminary processing is developed. In the new method, the uncertainty boundary for adversarial objects has become distinguishable and quickly computable. The hybrid method makes it possible to detect out-of-distribution (OOD) evasion attacks with a precision of not less than 80%, and SGM, MS, and BA attacks with a precision of 93%.